1# RECORD 84fe25f12d7833b71a83f2c2ae3a06dfc013ffcf2# TITLE: Do street-level scene perceptions affect housing prices in Chinese megacities? An analysis using open access datasets and deep learning3# YEAR: 2019.0 VENUE: PLoS ONE DOI: 10.1371/journal.pone.02175054# AUTHORS: Xiao Fu; Tian-Xia Jia; Xue-Qi Zhang; Shanlin Li; Yonglin Zhang5# PAGES: 18 CANDIDATE TABLE PAGES (1-based): [9, 10, 13]6# ABSTRACT: Many studies have explored the relationship between housing prices and environmental characteristics using the hedonic price model (HPM). However, few studies have deeply examined the impact of scene perception near residential units on housing prices. This article used house purchasing records from FANG.com and open access geolocation data (including massive street view pictures, point of interest (POI) data and road network data) and proposed a framework named “open-access-dataset-based hedonic price modeling (OADB-HPM)” for comprehensive analysis in Beijing and Shanghai, China. A state-of-the-art deep learning framework and massive Baidu street view panoramas were employed to visualize and quantify three major scene perception characteristics (greenery, sky and building view indexes, abbreviated GVI, SVI and BVI, respectively) at the street level. Then, the newly introduced scene perception characteristics were combined with other traditional characteristics in the HPM to calculate marginal prices, and the results for Beijing and Shanghai were explored and compared. The empirical results showed that the greenery and sky perceptual elements at the property level can significantly increase the housing price in Beijing (RMB 39,377 and 6011, respectively) and Shanghai (RMB 21,689 and 2763, respectively), indicating an objectively higher willingness by buyers to pay for houses that provide the ability to perceive natural elements in the surrounding environment. This study developed quantification tools to help decision makers and planners understand and analyze the interaction between residents and urban scene components.789===== PAGE 1 (score 5, numeric-paren 0, table-words 0, var-words 5) =====1011RESEARCH ARTICLE12Do street-level scene perceptions affect13housing prices in Chinese megacities? An14analysis using open access datasets and deep15learning16Xiao Fu1, Tianxia Jia1,2, Xueqi Zhang1,2, Shanlin Li3, Yonglin ZhangID1,2*171 State Key Laboratory of Urban and Regional Ecology, Research Center for Eco-Environmental Sciences,18Chinese Academy of Sciences, Beijing, China, 2 College of Resources and Environment, University of19Chinese Academy of Sciences, Beijing, China, 3 National Science Library, Chinese Academy of Sciences,20Beijing, China21* ylzhang_rcees@foxmail.com22Abstract23Many studies have explored the relationship between housing prices and environmental24characteristics using the hedonic price model (HPM). However, few studies have deeply25examined the impact of scene perception near residential units on housing prices. This arti-26cle used house purchasing records from FANG.com and open access geolocation data27(including massive street view pictures, point of interest (POI) data and road network data)28and proposed a framework named “open-access-dataset-based hedonic price modeling29(OADB-HPM)” for comprehensive analysis in Beijing and Shanghai, China. A state-of-the-30art deep learning framework and massive Baidu street view panoramas were employed to31visualize and quantify three major scene perception characteristics (greenery, sky and build-32ing view indexes, abbreviated GVI, SVI and BVI, respectively) at the street level. Then, the33newly introduced scene perception characteristics were combined with other traditional34characteristics in the HPM to calculate marginal prices, and the results for Beijing and35Shanghai were explored and compared. The empirical results showed that the greenery36and sky perceptual elements at the property level can significantly increase the housing37price in Beijing (RMB 39,377 and 6011, respectively) and Shanghai (RMB 21,689 and 2763,38respectively), indicating an objectively higher willingness by buyers to pay for houses that39provide the ability to perceive natural elements in the surrounding environment. This study40developed quantification tools to help decision makers and planners understand and ana-41lyze the interaction between residents and urban scene components.42Introduction43Urban scenes can be considered a vital medium for human beings, and it is important to44recognize and understand this aspect of cities [1] to ensure the spatial security of urban45sustainability [2]. The spatial distributions of urban scenes exhibit variety, complexity and46PLOS ONE | https://doi.org/10.1371/journal.pone.021750547May 30, 2019481 / 1849a111111111150a111111111151a111111111152a111111111153a111111111154OPEN ACCESS55Citation: Fu X, Jia T, Zhang X, Li S, Zhang Y (2019)56Do street-level scene perceptions affect housing57prices in Chinese megacities? An analysis using58open access datasets and deep learning. PLoS59ONE 14(5): e0217505. https://doi.org/10.1371/60journal.pone.021750561Editor: Shihe Fu, Xiamen University, CHINA62Received: January 16, 201963Accepted: May 12, 201964Published: May 30, 201965Copyright: © 2019 Fu et al. This is an open access66article distributed under the terms of the Creative67Commons Attribution License, which permits68unrestricted use, distribution, and reproduction in69any medium, provided the original author and70source are credited.71Data Availability Statement: All relevant data are72within the manuscript and its Supporting73Information files.74Funding: All authors were funded by The National75Key Research and Development Program of China76(2016YFC0503605). The funders had no role in77study design, data collection and analysis, decision78to publish, or preparation of the manuscript.79Competing interests: The authors have declared80that no competing interests exist.818283===== PAGE 9 (score 31, numeric-paren 0, table-words 3, var-words 15) =====8485descriptions of the variables are shown in Table 1. The data frames used to build the HPMs86can be found at https://doi.org/10.6084/m9.figshare.7823975.v1.87The independent variables in this article were classified into four major categories: location,88structure, neighborhood and scene perception characteristics. The location characteristic89included one variable named CENTER, which represented the shortest road distance from a90residential unit to the city center. To be more realistic, road distance and road network buffers91were adopted for all distance calculations in this paper.92The structure characteristics included AREA (average usable area in the apartment), year93(year of construction), HS (households), FR (floor-area ratio), PF (property management fee),94GR (greening rate) and the dummy variables ORI and BU. The dummy variable ORI repre-95sented the main orientation of the buildings in the residential unit: if the buildings mainly face96south, the value was set to 1; otherwise, it was set to 0. The dummy variable BU represented97the major building type: if the building was a slab, the value was set to 1; otherwise, it was set98to 0.99Table 1. Variable descriptions and basic statistical information.100Cities101Beijing102Shanghai103Variables104Description105Mean Standard deviation Mean Standard deviation106Dependent variable107LNPRICE108Log selling price in 10,000 RMB (Chinese currency, US $1 = RMB 6.71)1095.641100.541115.431120.58113Location characteristic114CENTER115Road distance to city center (km)11610.871176.3211815.111199.30120Structure characteristics121AREA122Average usable area in the apartment (m2)12388.4512496.8412586.5812644.97127AGE1282018 minus the year of construction of the building12919.6413013.7513122.7813223.90133ORI134Dummy variable, 1 if the building has windows facing south1350.781360.411370.931380.25139HS140Number of households141103014212001431037144942145FR146Floor-area ratio1472.521481.551492.271501.61151PF152Property management fee (RMB/m2 per month)1531.761541.421551.271561.28157GR158Green coverage rate (%)1590.321600.071610.341620.11163BU164Dummy variable, 1 if the building is a slab; 0 if the building is any other165type1660.601670.491680.871690.34170Neighborhood characteristics171AIRP172Road distance to the nearest airport (km)17322.131746.7517513.561768.58177BUS178Road distance to the nearest bus station (km)1790.241800.201810.181820.16183SUB184Road distance to the nearest subway station entrance (km)1851.471861.331871.631882.11189TRAIN190Road distance to the nearest train station (km)1914.741923.041937.091945.26195FINAN196Road distance to the nearest financial facility (km)1970.261980.361990.322000.36201RESTA202Road distance to the nearest restaurant (km)2030.142040.212050.152060.22207HOSP208Road distance to the nearest hospital (km)2090.182100.222110.232120.24213EDU214Road distance to the nearest educational facility (km)2150.182160.242170.212180.24219SHOP220Road distance to the nearest shopping mall (km)2210.102220.172230.092240.16225PARK226Road distance to the nearest park (km)2271.662281.342291.462301.08231WATER232Road distance to the nearest water body (km)2330.732340.502350.352360.28237WATER_A238Area of the nearest river or lake (km2)2390.122400.222410.112420.28243Scene perception244characteristics245LNGVI246Logarithmic mean green view index within 800 m road distance2472.802480.342492.972500.31251LNBVI252Logarithmic mean building view index within 800 m road distance2532.222540.542552.372560.55257LNSVI258Logarithmic mean sky view index within 800 m road distance2593.742600.162613.652620.26263https://doi.org/10.1371/journal.pone.0217505.t001264Street-level scene perceptions affect housing prices in typical Chinese megacities265PLOS ONE | https://doi.org/10.1371/journal.pone.0217505266May 30, 20192679 / 18268269270===== PAGE 10 (score 26, numeric-paren 0, table-words 4, var-words 15) =====271272The neighborhood characteristics included the road distances between the housing estates273and diverse urban amenities and environmental infrastructures, including the shortest road274distances to airports (AIRP), bus stations (BUS), subway stations (SUB), train stations275(TRAIN), financial facilities (FINAN), restaurants (RESTA), medical facilities (HOSP), educa-276tional facilities (EDU), commercial facilities (SHOP), urban parks (PARK), and water bodies277(WATER), as well as the area of the nearest water body (WATER_A).278Finally, the newly introduced scene perception characteristics (GVI, BVI and SVI) at the279residential unit level were transformed by the natural logarithm before building the HPMs.280The coefficients of the independent variables can explain the percentage change in the depen-281dent variables in the HPMs.282Results and discussion283Mapping scene perception indicators284To improve the observations and comparisons, this study plotted the street-level GVI, BVI and285SVI values in thematic maps for Beijing and Shanghai in 2015, as shown in Fig 5. The three view286indexes (VIs) were each mapped using five value intervals, and the value intervals in Beijing and287Shanghai were the same. The GVI value intervals were [0.00, 0.10], [0.11, 0.20], [0.21, 0.30], [0.31,2880.40], and [0.41, 0.70]; the BVI intervals were [0.00, 0.05], [0.06, 0.10], [0.11, 0.20], [0.21, 0.30], and289[0.31, 0.65]; and the SVI intervals were [0.00, 0.20], [0.21, 0.35], [0.36, 0.45], [0.46, 0.55], and [0.56,2900.65], which correspond to the qualitative levels of very low, low, medium, high and very high.291The statistical results indicate that the average GVI values in Beijing and Shanghai are 16.44292and 19.49, respectively, which indicated that pedestrians in Shanghai tend to see more green293scenes than those in Beijing. Fig 5(A) and 5(B) indicate that high and very high GVI road seg-294ments have a broader distribution and in general look “greener” in Shanghai than in Beijing.295The GVI values of road segments above the high level are mainly located in the central areas296(Dongcheng and Xicheng Districts), the northwestern area (Haidian District) and part of the297southern area, but the high GVI segments are mainly concentrated along the western bank of298the Huangpu River in Shanghai. The statistical results indicate that the average BVI value in299Shanghai is slightly higher than that in Beijing (10.7 and 9.21, respectively). The spatial distri-300butions in Fig 5(C) and 5(D) show that the BVI values in both cities present low-inside-and-301high-outside gradient variations. The spatial distributions of the SVI values in Beijing and302Shanghai show the opposite pattern relative to the BVI values (Fig 5(E) and 5(F)) and present303high-inside-and-low-outside gradient variations. According to the statistical results, the aver-304age SVI of Beijing is higher than that of Shanghai (42.10 and 38.47, respectively). Therefore,305from an overall perspective, people in Shanghai tend to see more profile areas of greenery and306buildings, whereas people in Beijing tend to see more open road canyons.307OLS regression modeling308The OLS regression modeling was four-fold: first, all independent variables were incorporated309into the OLS model calculation; second, all VIs were removed from the model, but other vari-310ables remained the same; third, the GVI was added to the OLS model; fourth, the GVI and311BVI were added to the model. The last three models controlled the VIs to show their contribu-312tions to the adjusted R2, and we included the last three models in the Supporting Information313to show the soundness of our model (S1 File).314The model results in Table 2 show that the adjusted R2 values were 0.7242 and 0.7511 for Bei-315jing and Shanghai, respectively, and most of the independent variables were statistically316significant. The variance inflation factor (VIF) test was performed to verify whether the variables317exhibited multicollinearity. The VIFs of the variables (less than 4) indicated that the318Street-level scene perceptions affect housing prices in typical Chinese megacities319PLOS ONE | https://doi.org/10.1371/journal.pone.0217505320May 30, 201932110 / 18322323324===== PAGE 11 (score 0, numeric-paren 0, table-words 0, var-words 0) =====325326Fig 5. The spatial distributions of VIs in Beijing and Shanghai (GVI, BVI and SVI).327https://doi.org/10.1371/journal.pone.0217505.g005328Street-level scene perceptions affect housing prices in typical Chinese megacities329PLOS ONE | https://doi.org/10.1371/journal.pone.0217505330May 30, 201933111 / 18332333334===== PAGE 13 (score 40, numeric-paren 0, table-words 4, var-words 16) =====335336housing prices through the newly built OADB-HPM framework. The model results in the pre-337vious section showed that street-level GVI and SVI had positive effects on the neighboring338housing prices in both Beijing and Shanghai, indicating that house buyers may have require-339ments for urban greenery and sky openness; however, BVI was not significant in any model,340indicating that house buyers did not favor more visible artificial buildings in their living envi-341ronment. According to the marginal price calculation, an increase in GVI and SVI by 1% can342increase housing prices by 0.2273% and 0.0899%, respectively, in Beijing and by 0.1853% and3430.0466%, respectively, in Shanghai (the conversion formula for the marginal prices for the344Table 2. OLS regression model (all variables).345Model 1 (Beijing)346Model 2 (Shanghai)347Variables348Full name349Unstandardized350Coefficients351Standard Error352VIF353Unstandardized354Coefficients355Standard Error356VIF357Constant358Constant3593.8283600.41163615.553620.2933363Location characteristic364CENTER365City center366-0.04523670.00123682.51369-0.01543700.00123713.45372Structure characteristics373AREA374Housing area3750.00793760.00023771.553780.00993790.00013801.39381AGE382Building age383-0.00273840.00013851.19386-0.00013870.00013881.05389ORI390Orientation3910.06793920.01313931.133940.05323950.02043961.02397HS398Households3990.00004000.00004011.364020.00004030.00004041.49405FR406Floor-area ratio407-0.00844080.00344091.184100.00004110.00004121.01413PF414Property management fee4150.03244160.00544171.474180.00214190.00284201.13421GR422Green coverage rate4230.24214240.08224251.184260.27444270.08324281.35429BU430Building types4310.03224320.01194331.214340.01894350.02254361.05437Neighborhood characteristic438AIRP439Airport4400.01764410.00914421.17443-0.00774440.00134453.46446BUS447Bus station448-0.01174490.03154501.15451-0.04474520.03924531.14454SUB455Subway station456-0.01724570.00484581.59459-0.00734600.00394611.60462TRAIN463Train station4640.00244650.00224661.83467-0.00064680.00134691.88470FINAN471Financial facility472-0.08584730.01894741.94475-0.08924760.01904771.78478RESTA479Restaurant4800.10174810.03894822.184830.04214840.05114851.65486HOSP487Hospital4880.01514890.03604901.904910.01164920.041344931.85494EDU495Educational facility496-0.06144970.03304981.46499-0.09585000.03625011.69502SHOP503Shopping mall5040.11495050.051785062.695070.10425080.06775092.70510PARK511Park512-0.00335130.00485141.62515-0.01535160.00595171.40518WATER519Water body5200.03375210.01025221.165230.04165240.02465251.22526WATER_A527Water body area5280.04475290.02155301.125310.02265320.03675331.08534Scene perception characteristics535LNGVI536Green view index5370.22735380.08185393.315400.18535410.04715422.71543LNBVI544Building view index545-0.00055460.00185473.96548-0.00085490.00145502.64551LNSVI552Sky view index5530.08995540.02805553.175560.04665570.02985583.50559F ratio560200.5200561171.1500562Adjusted R25630.72425640.7511565Durbin-Watson5661.76285671.8212568 Indicates significance at the 10% level569 indicates significance at the 5% level570 indicates significance at the 1% level.571https://doi.org/10.1371/journal.pone.0217505.t002572Street-level scene perceptions affect housing prices in typical Chinese megacities573PLOS ONE | https://doi.org/10.1371/journal.pone.0217505574May 30, 201957513 / 18576577578===== PAGE 14 (score 17, numeric-paren 0, table-words 1, var-words 15) =====579580dummy logarithm transformed variables were ex−1, where x is the nonnormalized coefficient581in the model). Therefore, the house buyers in Beijing and Shanghai favor the perception of nat-582ural aspects (green vegetation and open sky).583Through a comparison of the model results, home buyers in Beijing and Shanghai exhibited584the same willingness to pay for the GVI and SVI, but there were significant differences in the585marginal prices. Residents in Beijing showed significantly more willing to pay for GVI than586those in Shanghai, which may be caused by the locations, climates and seasons of the two cities.587Moreover, the two cities were located at different latitudes, and the green leaf period in Beijing588is shorter than that in Shanghai due to climate, humidity and seasonal changes. This difference589may be one of the major reasons affecting people’s perception and willingness to pay for the590greenery around housing estates.591In summary, according to the results in this article, the government should attach impor-592tance to the adjustment and optimization of the visual scenery around residential areas and593high-rise building clusters to improve people’s visual access to natural elements. For example,594the construction of open spaces and green parks in the public spaces between neighboring595housing estates may help to meet the needs of people. The government can fine-tune and596reshape the urban spatial structure according to the needs of residents to build an appropriate597living environment to meet people’s practical needs. For instance, the configuration of three-598dimensional green walls, corridors and roads can increase the daily interaction opportunities599between people and green vegetation, and the appropriate use of domes made of transparent600material and open walls may increase the sense of openness in various spaces.601The government should pay attention to the potential economic impacts of urban scene602perception on real estate. On the one hand, in the absence of government control, real estate603developers could exploit urban green and open spaces for additional benefits. Over time, the604excessive accumulation of such benefits may lead to low-income consumers being unable to605bear the financial costs resulting from increased housing prices. This phenomenon is detri-606mental to the sustainable development of the environment, society and economy. This phe-607nomenon not only violates the fairness of the living environment in cities but also may harm608the overall interests of citizens. On the other hand, in terms of the pleasant visual environment609surrounding a property, developers almost unconditionally obtain the economic benefits of a610pleasant residential environment but do not contribute to the maintenance or optimization of611the urban visual landscape. Therefore, the government should collect real estate tax from real612estate developers according to the constructions of urban scenes, and this revenue could be613used to fund the optimization and maintenance of urban scenes. The government needs to614strengthen the supervision of housing sales in the real estate market, restrain behaviors that615abuse green scenes and open spaces as bargaining chips in the process of real estate sales to dis-616rupt the market, and prevent developers from maliciously raising housing prices by citing the617appropriateness of the public visual environment during the sales process. The goals of real618estate developers should be to properly use the influence of urban visual elements on housing619price on the basis of reasonable bidding and scientific planning, improve the competitiveness620of their real estate brands, and participate in the construction of high-quality residential621environment.622In summary, this article identified the actual requirements for visible greenery and sky623openness in Beijing and Shanghai based on the newly constructed OADB-HPM framework.624Due to the previous lack of technology and appropriate research framework, the potential625effect of the visible element characteristics of residential environments on housing prices has626often been ignored by researchers and decision makers. However, the framework and techni-627cal means proposed in this paper can now be used as an effective supplement and quantifica-628tion tool.629Street-level scene perceptions affect housing prices in typical Chinese megacities630PLOS ONE | https://doi.org/10.1371/journal.pone.0217505631May 30, 201963214 / 18633